đź’ˇ Inspiration
Modern enterprise infrastructures handle massive data workloads that push x86 architectures to their thermal and power limits. Traditional system governors operate on rigid, reactive thresholds, leading to micro-architectural thrashing, core latency, and resource starvation.
Inspired by Stephen Davies' Blueprints and David Tuffley's Beyond Coding, we designed HyperOpt-Agent. We asked ourselves a radical question: Can we build a truly autonomous, zero-trust system administrator that lives entirely in the background, handles the heavy lifting of high-frequency hardware datasets, and uses calculus and local AI to heal and optimize the operating system asynchronously?
⚙️ What it does
HyperOpt-Agent is a next-generation autonomous background agent tailored for the Intel Core Ultra 7 architecture. It operates 100% offline, running non-blocking background loops to poll low-level system metrics (Intel RAPL power registers and POSIX kernel logs).
Every 5 seconds, the agent executes a complex multi-paradigm asynchronous workflow:
- Calculus Integration: It computes a discrete Riemann Definite Integral using the trapezoidal rule over the rolling power buffer \( W(t) \) to monitor precise cumulative energy expenditure in Joules: $$ Joules = \int_{t_0}^{t_1} W(t) \, dt \approx \sum_{i=1}^{n-1} \frac{W(t_i) + W(t_{i+1})}{2} \Delta t $$
- Perl-Style Regex Scraping: It tokenizes high-throughput raw
dmesglogs via compiled regular expressions to isolate silent hardware faults (e.g., subsystem driver bottlenecks likeiwlwifi). - Cryptographic Integrity: It signs every telemetry observation row with an HMAC-SHA256 digital signature to secure the historical database (
telemetry_history.csv) against data injection or malicious tampering (inspired by Mike Rosulek's The Joy of Cryptography principles). - Cognitive Reasoning: It orchestrates Google Genkit and Ollama (Llama 3.1) on the local Intel NPU via IPEX-LLM to evaluate the state space and recommend an optimal CPU core affinity matrix (
taskset). - Ada Guardrails: Before updating the Linux Kernel Scheduler, an Ada-style strict runtime verification layer checks bounds and states to prevent OS deadlocks, falling back to safe deterministic states if the local LLM experiences an hallucination.
🛠️ How we built it
The core engine is built using Python 3.12 inside isolated containerized environments (Docker Compose & Production-Ready Kubernetes Deployment manifests).
- The ingestion layer utilizes Go and Erlang-style asynchronous concurrent mesh patterns (
asyncio.gather) to pull telemetry without blocking the main worker loop. - The local AI execution stack targets the dedicated Intel NPU hardware via the Intel IPEX-LLM library, integrating natively with Google Genkit's pluggable architecture.
- Data engineering steps leverage Hadley Wickham's Tidy Data layout principles using Pandas and NumPy data structures.
🛑 Challenges we faced
Deploying Google Genkit completely offline while decoupling it from default Cloud API quotas presented a steep integration barrier. Default packages continuously attempted online validation handshakes, leading to 429 Quota Exceeded constraints on cloud sandboxes. We achieved a breakthrough by refactoring the instantiation mechanics—stripping remote cloud extensions out of the active runtime execution path and mapping Genkit directly to local Ollama host endpoints, enforcing safe heuristic fallbacks.
📚 What we learned
We gained deep systems-engineering insights into the execution boundaries of locally accelerated LLMs on modern NPUs. We learned how to transition from traditional rigid rule-based monitoring tools into elastic, cognitive self-healing frameworks capable of autonomously repairing low-level POSIX scheduling topologies.
Built With
- asyncio
- autonomous-agents
- background-workers
- cryptography
- data-science
- docker-compose
- google-genkit
- hmac-sha256
- intel-npu
- ipex-llm
- kubernetes
- llama3.1
- numpy
- ollama
- pandas
- posix-kernel
- python
- riemann-integral
- self-healing
Log in or sign up for Devpost to join the conversation.